Mobile NLP ONNX packages
Version: gliner25-stanza115-v1.
Public downloadable ONNX conversions of pretrained GLiNER2.5 multilingual and Stanza 1.15.0 models. No new training was performed. Download the shared GLiNER package and only the Stanza language(s) needed. These packages do not include an LLM.
| Package | Download size | File |
|---|---|---|
| gliner | 516.4 MB | gliner-v1.zip |
| stanza-es | 391.1 MB | stanza-es-v1.zip |
| stanza-en | 402.5 MB | stanza-en-v1.zip |
| stanza-de | 397.2 MB | stanza-de-v1.zip |
| stanza-fr | 401.3 MB | stanza-fr-v1.zip |
| stanza-it | 402.5 MB | stanza-it-v1.zip |
| stanza-nl | 386.3 MB | stanza-nl-v1.zip |
| stanza-pt | 388.0 MB | stanza-pt-v1.zip |
Integrity and installation
release.json records the archive size, SHA-256 and individual model file hashes. Verify the archive hash, extract into a temporary directory, verify files against the included manifest.json, then atomically activate the directory. Preserve adjacent external ONNX data files. Cache packages locally. Pin downloads to a specific Hub commit rather than mutable main in production.
Runtime compatibility
These exports are custom runtime fragments, not drop-in AutoModel or stanza.Pipeline checkpoints. A matching tokenizer, tensor preparation and decoding implementation is required. The GLiNER encoder uses quantized vocabulary weights and FP32 components, with boundary heads for lengths 64/128/256; candidate-valid outputs use UINT8 for mobile ONNX Runtime compatibility. Stanza packages include tokenization, POS/morphology, dependency parsing, character models and language-dependent lemma/MWT components and metadata. This release distributes model assets, not an inference service or a complete runtime SDK.
Sources and modifications
- Fastino GLiNER2.5 multi, revision
cf5593a5d45e3bbf204b9df621b13b1c0cf25ee3. - Stanford NLP Stanza, runtime 1.15.0, pretrained language repositories
stanfordnlp/stanza-{es,en,de,fr,it,nl,pt}.
Modifications are ONNX conversion, mobile packaging, GLiNER vocabulary quantization and output compatibility casts; no task-specific fine-tuning. Upstream repositories declare Apache-2.0; see LICENSE and NOTICE.
Limitations
Entity boundaries/types and dependency parses can be wrong. These assets do not by themselves guarantee reliable semantic relations, negation, or question handling. Mobile performance depends on device, runtime and memory pressure. No universal seven-language accuracy claim is made.